資料淨化於類別不平衡問題: 機器學習觀點

NCUIR > School of Management at National Central University > Executive Master of Information Management > Electronic Thesis & Dissertation > Item 987654321/81282

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/81282

Title:	資料淨化於類別不平衡問題: 機器學習觀點
Authors:	蔡武霖;TSAI, WU-LIN
Contributors:	資訊管理學系在職專班
Keywords:	機器學習;資料探勘;類別不平衡;抽樣;特徵選取;Machine Learning;Data Mining;Class Imbalanced Problem;Sampling;Feature Selection
Date:	2019-07-03
Issue Date:	2019-09-03 15:42:35 (UTC+8)
Publisher:	國立中央大學
Abstract:	機器學習在Google Alpha Go出現之後再次受到矚目，這也顯現出收集資料的重要性。但在現實生活中，資料收集時的困難與限制會造成收集資料的不平均。這容易使得分類困難與不準確，因為特徵選取與不平衡處理(抽樣)後，影響分類器在向量空間中的學習與分類效能。本研究使用知名公開網站的資料集，並設計二個流程來發掘類別不平衡問題，而特徵選取與抽樣誰該放置於前或後，使用五種不平衡處理抽樣模組，分別為三增加少數抽樣法、二減少多數抽樣法放置於前後，另外特徵選取使用二種模組，並加入有無正規化在這二項流程上。分類器目前在類別不平衡中，最常被使用支持向量機(SVM)與決策樹(Decision Tree Classifier)的二種分類器進行分類。從本研究實驗過程得知，類別不平衡資料在先執行特徵選取之後，再執行不平衡處理(抽樣)，低資料量在抽樣後為 SMOTE 增加少數抽樣法，高資料量在抽樣後Random為減少多數抽樣法(Under Sampling)，特徵選取小於20建議使用PCA，20維以上使用GA，分類器SVM 為佳的分類器，至於資料是否要正規化為決策樹不使用、支持向量機使用。;After the invention of Alpha Go, machine learning caught the public eye and showed us the essential need for data collection. Nevertheless, in reality, data collection is often uneven owing to its many difficulties and confinement. Feature selection and imbalanced (Sampling) have inherent impacts on Classifier in vector space. This in turn impacts the ability of learning and classification which also leads to difficulty and inaccuracy during data classification. This research aims to utilize data from public websites to design two processes to excavate imbalanced (Sampling), feature selection and place sampling in the beginning and at the end. It will utilize five examples of imbalanced (sampling); three examples of increased over sampling and two of reduced under sampling placed in the beginning and the back. Moreover, it will use two different models and utilize normalization with non-normalization in the two processes. Classifier in class imbalanced is often used to support vector machines and decision trees two model. From this research, we can find out that class imbalanced need use after then use feature selection, SMOTE is when low data amounts after sampling increase over sampling. Random is when high data amounts after sampling reduce under sampling. It is recommended to use PCA when feature selection is under 20 dimensions, as GA is recommended if feature selection is above 20 dimensions. Moreover, the ideal classifier is SVM. When it comes to the question of utilizing normalization in data, we can utilize classification to selection. decision tree abandons it. support vector machines use it.
Appears in Collections:	[Executive Master of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	200	View/Open

社群 sharing

Loading...